Adaptive challenge cost: price by suspicion, on the clock not the hash

[cport1]

Difficulty was effectively constant — 4, or 5 from a datacenter address, or up to 6 under load. That is the wrong knob to escalate.

Browser JS runs 1–3M hash/s; native code 100–500M/s. Difficulty 6 costs a native solver about a millisecond and a budget Android phone about sixteen seconds. Raising it taxes the slowest legitimate devices and constrains nobody.

Wall-clock has no such asymmetry — nobody can make less time pass. Each challenge now carries a signed minAgeMs, and that is where escalation lands.

source	difficulty	minAgeMs
clean, or unknown	4	1500
1–2 recent strong bot verdicts	4	4000
3–5	5	8000
6+	5	15000

A clean visitor pays exactly what they paid before. At the 1.5s baseline one address mints roughly 40 tokens a minute; at 15s, four.

Difficulty now caps at 5 rather than 6, and a datacenter address no longer raises it at all — a real person on a corporate VPN or iCloud Private Relay was paying several seconds of blocked hashing for a shared IP. It raises the time floor instead, which a hosted scraper feels as reduced throughput and someone filling in a form does not feel at all.

Cost lands as latency, not as a score

The client is told minAgeMs and waits it out before submitting, so an ordinary visitor absorbs an elevated cost as a short delay rather than a worse verdict. That matters most for anyone sharing an egress address with whatever earned the delay.

A client that submits early is still scored, but the two cases are separated. Below the universal 1500ms baseline keeps its existing strength — nothing legitimate happens that fast. Between the baseline and the source's own floor is contributory only, because an older cached client does not know to wait and should not be read as automation for it.

Scope of the memory

Per (site key, address), 15-minute window, records only verdicts at or above 0.8, stores nothing but timestamps. Same shape and privacy class as the rate limiter beside it. It is not Workstream F §10.1 cross-session correlation — no behavioral vectors, no per-fingerprint history — and the module comment says so, because those two should not be conflated on the grounds that both are "server-side memory".

Measurement

Human FPR 0.00% (0/126), agent TPR 97.33% — both unchanged.

Two harness defects surfaced, both invisible until a signal was keyed by source.

Sample addresses collided. They were a hash of the sample id, putting 180 samples into a 508-address space. By the birthday bound: 25 shared addresses, 24 of them putting an agent and a human on the same IP. The corpus was manufacturing shared egress its labels never claimed, and the first source-keyed signal read 12.7% FPR on a panel whose real answer was zero. Addresses are now assigned by position, and the pool throws rather than wrapping.

Runs inherited each other's server state. One signal fired on 2/75 agents against a fresh server and 74/75 against one already benchmarked. Each run now namespaces its own site key.

The replayer deliberately still does not honour minAgeMs. A replayer that did would satisfy the timing gate by construction and the panel could never observe it firing on a human — the same trap as pinning a property during normalization and then reporting that nothing disagrees with it.

bench/package.json pointed npm test at lib/lib.test.js, which has never existed; CI runs npm test in server-node, so nothing invoked it. Replaced with real tests for the address pool and wired into the bench workflow.

Also in here

server-node/index.js was missed by the pass that removed HMAC truncation elsewhere: token signatures were cut to 64 bits on both generation and verification, while Go, Python and the standalone server used the full digest. Now full-length. Tokens issued by an older library build will not verify after upgrading.

Verification

• Go: go vet clean, all tests pass (10 new)
• Node: all suites pass (12 new); the new suite is wired into npm test, which it was not
• Python: 43 tests across 5 files pass (10 new)
• E2E: 103/103 (5 new)
• Bench harness: 5 new tests
• All three servers confirmed serving identical tiers end-to-end